IM-GRASP: Influencer-Guided Multimodal Graph for Cold-Start and Sparse Next Basket Recommendations in Social Commerce

In the rapidly evolving landscape of social commerce, platforms like TikTok, Instagram, and Twitter are reshaping consumer behavior through influencer-driven engagement and multimodal content exposure. Traditional Next Basket Recommender Systems (NBRS) face significant challenges in these environments, particularly under conditions of data sparsity and cold-start users or items. To overcome these limitations, we introduce a novel framework called Influencer-guided Multimodal Graph-based Recommendation with Adaptive Swarm Processing (IM-GRASP). First, we construct a heterogeneous multimodal graph that captures intra-entity (user-user, item-item), inter-entity (user-item), and cross-domain (influencer) couplings by integrating textual, visual, and behavioral signals. Second, we incorporate a biologically inspired Mass Spawning-based Particle Swarm Optimization (MS-PSO) mechanism to enhance representation learning and mitigate the over-smoothing problem commonly encountered in Graph Convolutional Networks (GCNs). Third, a lightweight contrastive intent modeling module is introduced to infer user intent sequences without relying on deep RNNs, thereby reducing sequential-model complexity and improving computational efficiency. Finally, we propose an engagement-weighted basket ranking mechanism that jointly considers user intent similarity and influencer-driven popularity. Extensive experiments on real-world datasets from TikTok, Instagram, and Twitter show that IM-GRASP provides competitive and consistently improved performance over the evaluated baselines in terms of recommendation accuracy, contextual relevance, novelty, and robustness in cold-start scenarios.